Compute capacity & supply
Commercial GPU capacity has buyer validation across multiple providers. Repeatable demand is established in parts of the market but remains uneven across the broader category.
ConfidenceLow
Assessed as of
Current state, informed by 36 months and relevant older history.
Assessment methodologySupporting sources (10)
- OpenAI agreement establishes a marquee model-provider motion
- Perplexity partnership centers inference workloads
- Signs multibillion-dollar Microsoft infrastructure agreement
- Poolside deployment used as flagship proof of rapid, managed cluster delivery
- Strengthened credibility with scale and strategic partnerships
- Published customer proof for dedicated model hosting with Reve AI
- Moreh customer proof showed inference performance gains on MI355X
- Expanded Microsoft GPU infrastructure deal
- Confidential AI deployment with ExpressVPN
- Cursor customer story
Providers with substantial commercial scale coexist with smaller participants whose adoption is less established. Competitive strength is unevenly distributed despite credible offerings across several independent providers.
ConfidenceLow
Supporting sources (11)
- OpenAI agreement establishes a marquee model-provider motion
- Perplexity partnership centers inference workloads
- Signs multibillion-dollar Microsoft infrastructure agreement
- Poolside deployment used as flagship proof of rapid, managed cluster delivery
- Strengthened credibility with scale and strategic partnerships
- Expanded Microsoft GPU infrastructure deal
- Anthropic signs ~$45B six-year compute deal with Nscale
- Reported contracted revenue reaches about $103B
- Figure commits $3.5B of compute in strategic Nscale partnership
- Moreh customer proof showed inference performance gains on MI355X
- Cursor customer story
Substantive change spans several independent providers. Access to compute capacity is broadening as production inference capabilities advance.
ConfidenceMedium
Supporting sources (12)
- CoreWeave expanded cross-cloud AI capabilities with Google Cloud
- Bare Metal Instances launch
- Bloom Energy power partnership
- Flash reaches GA
- TokenRouter launches as a unified AI model access platform
- Launched XFRA distributed data center solution
- Arm AGI CPU added to next-generation infrastructure
- Confidential AI deployment with ExpressVPN
- Provisioned Throughput launched for reserved inference capacity
- General availability with self-serve API and free credit
- Anthropic signs ~$45B six-year compute deal with Nscale
- Figure commits $3.5B of compute in strategic Nscale partnership
Venture investment has increased across both segments and reaches multiple independent recipients. Large commitments account for a substantial share of the increase.
ConfidenceMedium
Supporting sources (18)
GPU services are broadening beyond capacity rental into managed AI execution. Inference capabilities are extending into additional application needs across independent providers.
ConfidenceMedium
Supporting sources (15)
- Serverless RL launched for AI agents
- CoreWeave Inference became a unified inference platform
- Bare Metal Instances launch
- Nebius Token Factory launch
- Anthropic custom data-center deployment announced
- Flash reaches GA
- TokenRouter launches as a unified AI model access platform
- Hyperbolic Organizations launched
- Dedicated Model Hosting launched
- Launched XFRA distributed data center solution
- Launch of sovereign LLM-as-a-Service offering with Siili
- Terraform provider for Verda infrastructure
- Real-time voice agent stack launched
- Dedicated Container Inference launched
- Provisioned Throughput launched for reserved inference capacity
Documented new participation is increasing the range of independent compute providers. There is no clear broad reduction in independent participation.
ConfidenceLow
| Company | Score | Movement | History |
|---|---|---|---|
| 92 | — | ||
| 88 | — | ||
| 87 | — | ||
| 84 | +2 | ||
| 81 | — | ||
| 74 | +16 | ||
| 59 | — | ||
| 58 | — | ||
| 55 | — | ||
| 52 | — | ||
| 34 | — | ||
| 24 | — | ||
| 22 | — | ||
| 12 | — |